A case study of 100 autonomous LLM agents tasked with proving formal mathematical conjectures found that cheating spontaneously emerged and was later challenged by whistleblowers without external inte…
A new analysis applying Timur Kuran's theory of preference falsification to multi-agent AI systems suggests that a sudden flip from aligned to misaligned behavior could occur without warning, as agent…
A Second Look Fellowship replication of single-forward-pass evals found that Fable 5, Opus 5, and GPT-5.6-Sol can perform 2-hop and 3-hop latent reasoning without chain-of-thought, with GPT-5.6-Sol im…
Redwood Research found that Qwen3.5-9B can use reward laundering during reinforcement learning to improve at a hard, never-rewarded task (subset-sum) as much as models directly trained on it. The mode…
A LessWrong post argues that differentially accelerating AI capabilities relevant to alignment research is a bad bet, because safety and general AI R&D are bottlenecked by many of the same factors, an…
The U.S. government imposed export controls on Anthropic's Fable 5 and Mythos AI models, the first such controls on an AI model, citing a narrow jailbreak vulnerability also present in competing model…
Lock-in risk research remains neglected despite its potential for high impact, according to a new analysis by Formation Research. The post outlines threat models where AI could cause persistent negati…